FLEX scores 0.99 as robots learn broader human preferences

By Marcus Bennett

FLEX averaged a normalized reward of 0.99 in the navigation scenario requiring all three operations, according to the MIT–New York University paper. The paper assigns 1 to optimal behavior and 0 to the worst-case trajectory. FLEX’s 0.99 exceeded 0.69 for add-only ALGAE and 0.83 for mask-only Mask-From-Exhaustive (MFE).

The unscaffolded baseline scored 0.78. MIT announced FLEX on November 9, 2026, ahead of the researchers’ November 2026 presentation at the Conference on Robot Learning. MIT and New York University researchers developed the Flexible Feature Editing Pipeline, which uses conversational corrections to revise the environmental features guiding a robot’s movements.

What problem does FLEX solve?

The paper describes a failure mode in language-guided robotics (control shaped by natural-language instructions). A coffee-carrying robot told not to spill on a laptop could add that object to its obstacle list without capturing why it matters.

FLEX instead infers a broader preference for avoiding electronics. The simulated arm then adjusts its path around both the laptop and a nearby phone, without a separate phone-specific instruction.

The paper treats the instruction as evidence about a preference rather than a complete rule. A phone becomes relevant through the inferred reason, not because the user issued another command.

FLEX separates feature specification, meaning the selection of measurable scene properties, from reward optimization, meaning the tuning of how those properties shape movement. Language corrections identify the properties that matter, while demonstrations determine how strongly they should influence a trajectory. The paper says the separation targets representations with missing features, irrelevant over-specification or abstraction at the wrong level.

The paper’s two-part design lets a correction alter what the robot notices without retuning every movement preference at once. Its staged pipeline revises the representation one decision at a time.

The paper says each correction passes through three ordered operations:

• ADD — The stage checks whether existing features explain the correction. If they do not, a language model proposes a missing feature and generates code to compute it.
• GENERALIZE — The stage replaces an overly specific feature with one tied to the inferred intent. In MIT’s bowling-ball example, a mug-specific constraint becomes avoidance of fragile objects.
• MASK — The stage removes features that are no longer relevant to the correction.

The paper places MASK last, making relevance decisions after any generalization. The experiments used GPT-4o, and MIT reports that the full three-step interaction takes about five seconds.

Andreea Bobu, who led the study and heads MIT’s Collaborative Learning and Autonomy Research Lab, tied the design to household use: “Everyone’s homes are different, and everyone cares about different things, so no robot comes out of the box knowing what matters to you.”

Bobu said: “What excites me about FLEX is that it can pick that up from a sentence or two instead of asking you to spell out every rule — and that’s the difference between a robot people put up with and one they actually want in their home.”

How strong were the simulation results?

The researchers tested FLEX in 9-by-9 navigation grids and simulated tabletop tasks with a seven-degree-of-freedom Franka arm, or seven movement axes. The tasks involved coffee, a lit candle and a bowling ball.

The arm scenarios used five-stage correction sequences and changed object arrangements between stages. Those changes could make a previously useful feature irrelevant or expose a hidden misalignment. Each stage paired a correction with changes in the tabletop arrangement, allowing the researchers to observe whether a learned constraint remained useful after the scene shifted.

Comparisons included add-only ALGAE, mask-only MFE and an unscaffolded baseline that asked the language model for a complete updated feature set in one query. The paper reports that FLEX achieved the highest average normalized reward in all seven navigation scenarios.

Under the paper’s setup, ALGAE tested feature addition alone, while MFE tested feature deletion alone. The unscaffolded baseline placed the complete editing request in one query. In the scenario requiring all three operations, FLEX averaged 0.99, compared with 0.69 for ALGAE, 0.83 for MFE and 0.78 for the unscaffolded baseline.

In the sequential arm simulations, the authors report that FLEX generally learned the target reward after one or two corrections and maintained performance on earlier and later configurations. The comparison methods required more corrections and degraded as irrelevant features accumulated or their representations overfit individual stages.

What did the human study show?

A separate in-person study involved 12 participants, each providing typed corrections across three five-stage blocks using FLEX, ALGAE and MFE. The researchers report that FLEX received the highest average rating on all five questionnaire dimensions. The three blocks each contained five stages, and participants supplied corrections by typing rather than speaking.

For intent understanding on a seven-point scale, averages were 6.3 for FLEX, 5.0 for ALGAE and 4.5 for MFE. The paper reports that FLEX’s differences from both baselines were statistically significant after adjustment for multiple comparisons.

The researchers say differences in ease of use and collaborativity were not statistically significant.

What remains unproven?

The human study did not run closed loop on a physical robot, where the machine would act on live feedback. Participants watched simulated trajectories because some baseline methods produced erratic arm motions.

The team separately combined their corrections with expert demonstrations collected on a physical Franka for offline learning, meaning training after data collection. The authors describe that hardware work as a feasibility check supporting the simulation study.

FLEX remains limited to quasi-static manipulation, or handling rigid objects with limited dynamic motion, whose relevant properties can be expressed in language. The authors say dynamic tasks and deformable objects may require additional modalities (other sensing or control information).

Language descriptions can cover many rigid-object properties, but the authors do not extend that claim to fast motion or objects that change shape. They identify both cases as requiring information beyond the current language-based formulation.

The authors also say language-model commonsense can accommodate widely shared preferences but may fail for personal or counterintuitive ones. A request to stay even farther from a laptop encodes preference strength, which the current formulation does not capture.

Preference strength is separate from direction. Saying “farther” indicates which way to move, but it also signals how much distance the user wants. The current formulation captures the former, according to the authors, not the latter.

The authors’ project page remains marked as a work in progress and lists code and dataset availability as coming soon. The open research question is whether language-model generalizations can accommodate idiosyncratic preferences with the same consistency during closed-loop operation on physical robots.

For household and assistive robots, fewer corrections help only when the inferred preference matches what the user actually cares about, the researchers say. Neither the 0.99 simulation result nor the 6.3 intent score settles that question, because no study tested FLEX in closed-loop physical control.

The authors frame the central next test as whether FLEX can retain that level of intent understanding while a physical robot applies an inferred preference during live operation.